Milind Borkar
Strategic Consultant, VICON
Although only a small number of airports are using video analytics in a sophisticated manner, there is no question that the technology offers the potential for improving video-based security operations.
The questions which the airport must address in considering whether and to what extent video analytics should be deployed are:
- What video analytic functions are needed and where?
- Can they be defined accurately and deployed properly?
- Will they produce the expected benefits including tolerable false alarm rates?
The false alarm issue is fundamental. A single or even a small number of cameras which false alarm several times a day may be tolerated, but San Diego International Airport (SDIA) is looking at a population of some 400 cameras through the AMP2 expansion program. Even if 10 percent, or 40 cameras, are set up with video analytics, and if all or most of alarms make false alarm several times each day, the alarms would overwhelm the SOC and the system would be shut down.
Exterior and interior opportunities for using video analytics at SDIA present different challenges, and have to be treated differently.
The video cameras now being used for exterior perimeter surveillance are not suitable for video analytics at night for a variety of reasons including:
- Inadequate scene illumination, and in some areas, strong point light sources which blind the cameras.
- Distances from target to camera which force lenses to be zoomed to maximum focal length, reducing the amount of light on the detector and degrading its performance (the so-called f/# problem).
- The inability of visible light cameras to operate in fog.
Better cameras and lenses, and better and more controlled scene lighting, may enable video analytics to be effective for perimeter surveillance, but that needs to be demonstrated. Infrared and thermal cameras may be better choices, as they eliminate the problems of scene illumination and poor visibility, but they, too, must be subject to testing.
The following images, recorded by Siemens Corporate Research, illustrate some of the scene variables that have to be dealt with in applying video analytics to exterior surveillance. Scene conditions at night, when target contrast is usually reduced and bright point light sources in the scene may be present (as in the cargo area of SDIA), can be especially challenging.
Applying video analytics indoors should present fewer problems. Scene illumination should be under the airport’s control, wind and fog will not be issues, distances will be relatively short etc. That does not, however, avoid the need for careful planning and testing, both of which are prerequisites to realizing effective performance.
The potential impact of false alarms from a large number of cameras argues for a minimalist approach in which a few mature video analytic functions are installed and then are subject to rigorous testing in the real airport environment before being extended.
Each analytic function will present its own unique requirements. This process starts with defining the minimum functionality needed at each camera site, which is much an airport responsibility as a technical issue; the environment in which the function must perform (camera motion being a critical problem, since video analytics are designed to work with fixed cameras, along with scene illumination); camera placement and aspect needed to achieve the performance objectives; and the maximum false alarm rate that can be tolerated.
In this context, being able to do one function very well is more important than doing several functions less well. Of the many offerings available from different vendors, the following list is representative of the video analytic functions that SDIA might consider adopting selectively based on the specific requirements of each situational condition:
- Video Motion Detection: Detect and track multiple moving objects in a Region of Interest (RoI) defined in the camera field-of-view. VMD is perhaps the most mature of the video analytic functions, and is now routinely bundled with video cameras. Generally not complicated to operate, but also does not provide sophisticated target discrimination.
- Tripwire: Detect when an object crosses a defined threshold or multiple thresholds. This is a variation of the VMD analytic, the RoI essentially being reduced to a line or set of lines.
- Counter Flow: Detect an object moving counter to normal flow direction e.g., on a jetway or along a fenced perimeter. Effective in uncluttered environments, especially indoors, complicated when confronted with dynamic clutter.
- Congestion: Alert the SOC when the number of objects in a Region of Interest exceeds a threshold. Evolved from VMD, this function is sensitive to camera position and requires more precise and complicated rules than VMD.
- Abandoned Object: Detect an object left unattended for a defined period of time. Perhaps one of the more complicated video analytic functions to implement because of the many possible target scenarios. Scenario variations include multiple objects at variable times, objects being occluded for.
Variable lengths of time by passenger traffic flow and other objects of variable size and dwell time, and persons who move in and out of the ROI set for associating them with the targeted objects.
The function is vulnerable to the presence of other objects of similar size in the camera field of view, target-to-background contrast, the movement of persons in front of and around the object that can grow quickly. Drawing a box (ROI) around an object is the easy part (and the object should occupy 25 percent or more of the field-of-view of the camera for best results).
Defining what constitutes an ‘object of interest’ worthy of setting an alarm may be based on criteria such as the amount of time since the object last moved and on rules established for object behavior.
Associating objects with owners can be very complicated, especially when other passengers are present and move in and out of the ROI. If a bag enters the terminal lobby with two persons and one or both leave the bag at different times, and one returns but other persons also appear in the ROI, the software will be challenged to associate ‘ownership’ with the proper person and establishing a rule for what constitutes an alarm will be similarly challenged. Heft Detection, Stationary Object Detection, and Loitering Object Detection are variations of the abandoned object function, each with its own peculiarities.
- Tailgating/ Piggybacking Detection: Detect one or more unauthorized persons moving through a secured portal being used by an authorized person. Reliable detection is sensitive to camera placement (two or more cameras may be needed to cover the persons from the necessary aspects), scene illumination (to realize full camera performance), speed of passage and the size of the gaps between persons, and the ability to discriminate between inanimate objects such as ladders, which are carried through the portal. Multiple cameras may be required to realize a satisfactory false alarm rate.
Video analytics come in several ‘flavors’ including policy-based, rules-based, and behavior-based – these terms being relatively loosely defined. The classes will often overlap, with each new class increasing detection performance but also computational complexity (and cost).
Behavioral analytics, for example, intelligently monitor all the standard detected features of moving objects and build up a concept, over a large period of time, of what motion can be deemed as typical, and thus can be ignored. Events then become triggered by abnormal behavior which the airport operator should define. A pedestrian loitering and/ or approaching a number of different cars in a parking lot in a short period of time, indicating potential criminal activity, could be viewed as abnormal behavior, and consequently generating an event.
How the video analytics function is less important than what they deliver for airport security in the real-word environment, whether they be understood by the operators, and whether they be adapted to the moves/ adds/ changes typical of an airport operation.
Video analytics are best applied to full resolution (4CIF) images at full frame rate (30 fps), because this presents the greatest amount of information in the shortest time for analysis. Proposals to apply video analytics to CIF images (quarter frame) and at reduced frame rates, typically 7 to 12 fps, are more susceptible to false alarms.
The best place to perform live video analytics is at the camera, at the edge of the network, where the analytics can be applied to the video signal before it is compressed or has its frame rate reduced to conserve transmission bandwidth. This also provides a reliability advantage, since a camera failure only impacts that camera whereas if a central server fails multiple cameras become inoperative.
The best place to locate post-event processing is on a central server, so that recorded video can be searched many times and using different parameters.
